Xinference Documentation

Xinference is a cloud platform for deploying and serving open-source AI models at scale. With a single API call you can launch large language models (LLMs) and embedding models on dedicated GPU instances — billed per second, no infrastructure management required.

What is Xinference?

Xinference provisions GPU-backed inference clusters on demand. When you create a deployment, the platform:

  1. Selects the optimal GPU instance type for your chosen model
  2. Provisions the EC2 cluster in the cloud
  3. Downloads the model weights (or restores them from cache)
  4. Starts the Xinference inference server
  5. Returns an OpenAI-compatible endpoint you can query immediately

When you terminate a deployment the cluster is torn down and billing stops.

Core Concepts

Concept Description
Deployable Model A model variant pre-approved for deployment (e.g. qwen2.5-instruct in int4 quantization on a specific GPU family).
Deployment A running instance of a deployable model assigned to your account, with its own endpoint URL.
Cluster The set of EC2 instances (supervisor + workers) backing a deployment.
Organization A group of users sharing billing and deployment quotas.
Billing Account Tracks usage and balance for an organization. Metered in GPU instance-seconds.

Key Features

  • OpenAI-compatible API — drop-in replacement for openai.ChatCompletion and openai.Embedding
  • Per-second billing — pay only for the time your model is running
  • Model caching — popular models cached on S3 for fast cold-start times
  • Multi-user organizations — invite teammates and share a single billing account
  • SSO support — sign in with Google
  • Idle auto-termination — clusters shut down automatically after a configurable idle period

Supported Model Types

  • LLM (Chat) — instruction-tuned language models for chat and text generation
  • Embedding — text embedding models for semantic search and RAG